analyze-correlations

name	analyze-correlations
description	Run correlation analysis on connected data, identify strong relationships, and suggest regression models. Use when exploring relationships between variables.
allowed-tools	mcp__localdata__describe_database mcp__localdata__execute_query mcp__localdata__analyze_hypothesis_test mcp__localdata__analyze_effect_sizes
argument-hint	<database-name>

Analyze Correlations

Discover and quantify relationships between variables in a connected dataset.

Steps

1. Get the schema. Call describe_database with the database name from $ARGUMENTS. Identify all numeric columns across tables. Note column names and which table each belongs to.

2. Extract numeric data. Call execute_query to select all numeric columns from the primary table. If data spans multiple tables, join on shared keys. Limit to 50,000 rows if the dataset is large.

3. Test pairwise correlations. For each meaningful pair of numeric columns, call analyze_hypothesis_test with test type "correlation". Use Pearson for normally distributed data, Spearman for skewed or ordinal data. Record the correlation coefficient and p-value for each pair.

4. Filter significant results. Retain pairs where p-value is below 0.05. Sort by absolute correlation strength. Classify relationships:

   • Strong: absolute r above 0.7
   • Moderate: absolute r between 0.4 and 0.7
   • Weak: absolute r between 0.2 and 0.4

5. Measure effect sizes. For the top 5 strongest correlations, call analyze_effect_sizes to quantify practical significance. Compare statistical significance (p-value) against practical significance (effect size). Flag cases where a correlation is statistically significant but practically negligible.

6. Check for confounders. Look for pairs of strong correlations that share a common variable. Note potential confounding relationships where A correlates with B and A correlates with C.

7. Present results. Provide a ranked table of correlations with columns: variable pair, correlation coefficient, p-value, effect size, and interpretation. Group by strength category.

8. Recommend next steps. For the strongest relationships:

   • Suggest /localdata-mcp:regression with the best predictor-outcome pairs
   • Flag multicollinearity risks if predictors are highly correlated with each other
   • Recommend further investigation for surprising or counterintuitive correlations